Record Matching (RM)finds out instances referring to the same entity between different data sources.Existing work mainly uses the similarity between the key attribute values of instances for RM,while seldom work employs non-key attribute values.As a result,when two instances referring to the same entity do not have similar key attribute values,they might be missed as a matching pair.On the other hand,some particular non-key attribute values shared by the two instances might reflect the relationship between them.Based on the intuition,we propose a novel RM method based on non-key attribute values.Compared to key attribute,non-key attributes can be more noisy and inconsistent.Besides,there are usually a lot more non-key attributes than key attributes,thus RM based on non-key attributes faces a significant efficiency problem.To deal with these challenges,we propose a rule-based algorithm based on a tree-like structure.With this tree-like structure,we can not only deal with noisy and missing values,but also greatly improve the efficiency of the method by finding out matched instances or filtering unmatched instances as early as possible.The experimental results based on several data sets demonstrate that our method outperforms existing RM methods by reaching a higher precision and recall.Besides, the proposed techniques can greatly improve the efficiency of a baseline algorithm beyond 10 times.%实体匹配旨在找出不同数据源中指代同一实体的实例。已有的实体匹配方法大都基于实体主属性值的相似度进行匹配，而很少有工作考虑到使用实体的非主属性值来辅助实体匹配。然而，当两条指代同一实体的主属性值差异较大的时候，这两个实体可能不会被认为是匹配的实体。另一方面，这两个实体很可能共享一些特别的非主属性值，而这些非主属性值恰好可以反映出两个实体的匹配关系。基于这种思想，文中提出了一种新颖的基于非主属性值的实体匹配算法。该算法以类似于决策树的结构为基础，通过使用这种结构，不仅可以解决噪声值和空缺值带来的问题，而且可以极大地提高发现匹配记录以及尽可能早地排除不匹配记录的效率。多个数据集上的实验结果表明我们的方法比现有的实体匹配方法具有更高的准确率和召回率。此外，使用我们提出的基于决策树的匹配算法等有关技术较 Baseline 匹配算法在匹配效率上高出10倍多。
展开▼
